EXPLAINABLE MACHINE LEARNING FOR EVALUATING DIABETES PREDICTION MODELS

Abstract

The present study has investigated the use of Explainable AI (XAI) machine learning models for diabetes prediction, with a particular focus on SHAP (Shapley Additive Explanations) values and  XGBoost importances (gain). Using the publicly available Pima Indians Diabetes dataset there were various classification models: Logistic Regression, Random Forest, and XGBoost trained and evaluated based on the parameters: Accuracy, and AUC. Global and local explainability was further assessed through SHAP bar plots and XGboost gain importances plots were used for analyzing structural feature importances. Results gathered suggested that stabilized glucose level and age were the most influential predictors of diabetes across models even though differences emerged in the ranking of the feature HDL (High-Density Lipoprotein) cholesterol. Overall, while SHAP provided a hybrid understanding of local and global patient-specific explanations, XGboost gain importances gave structural importance. The discussion section also includes how the combination of these two tools provided a hybrid, comprehensive and nuanced understanding about how these models bridge the “black-box” gap in clinical decision-making, ensuring trust, clarity and transparency in the prediction of AI models for sensitive domains like healthcare applications.